A1281
Title: A regularized regression approach to global minimum variance allocation
Authors: Timm Pfeil - University of Innsbruck (Austria) [presenting]
Dennis Umlandt - University of Innsbruck (Austria)
Abstract: A regularized regression approach for constructing the global minimum variance portfolio (GMVP) is proposed in large equity universes. Rather than shrinking the return covariance matrix, allocation weights formulated as regression coefficients are directly penalized. This regression-based formulation reduces the number of parameters to be estimated and ensures well-posed solutions even when the number of assets exceeds the sample length ($N > T$). The method is applied to estimate GMVP weights for S\&P 500 constituents from 1957 to 2022. Relative to static and dynamic benchmark approaches, the penalized allocations consistently reduce realized out-of-sample variance and improve Sharpe ratios while maintaining comparable or higher average returns. Lasso and elastic net produce sparse, low-turnover allocations, whereas ridge yields stable, diversified exposures. The results are robust across window lengths, index constructions, and crisis subsamples, while preserving the GMVP interpretation, in which regression coefficients map directly to portfolio weights.
